acceptodds
Under review as a conference paper at ICLR 2027

SSDenoDet: Self-Supervised Cross-Domain Feature Denoising for SAR Target Detection

Abstract

Labeled synthetic aperture radar (SAR) detection data are costly, while speckle and small targets challenge conventional detectors. We present SSDenoDet, which transfers a corruption-robust masked-autoencoding representation learned from hyperspectral imagery (HSI) to SAR. A frozen MAESSI encoder and centered exponential-moving-average targets train Yield-Aware MAESSI-Aligned Learner (YAMAL), a grayscale SAR image encoder; an FPN then localizes targets from its spatial features. HSI and the teacher are absent at inference. SSDenoDet-E reaches 57.55 COCO mAP on untouched SARDet-100K validation versus 55.60 for our matched DenoDet V2 reproduction, and improves common-evaluator AP50 over its single-model counterpart on four independently trained SAR benchmarks.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Deep learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_dEZnOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes SSDenoDet, a framework for transferring representations learned from hyperspectral imagery (HSI) to SAR object detection without paired HSI–SAR scenes. A frozen, corruption-pretrained MAESSI encoder supervises a ResNet-50 student through local and global alignment on grayscale HSI projections. In parallel, an EMA teacher provides self-distillation targets on augmented SAR images. The resulting backbone is fine-tuned with an FPN-based supervised detector; HSI and teachers are unnecessary at inference.

The claimed contributions are the HSI-to-SAR robustness-transfer formulation, the combination of spatially aligned cross-domain supervision with SAR consistency and anti-collapse regularization, and a detection pipeline that preserves small-target scale. SSDenoDet-E additionally uses multiple checkpoints or flipped views to aggregate predictions.

Validation covers SARDet-100K, AIR-SARShip-1.0, SAR-AIRcraft-1.0, and SARBench, a split derived from HRSID. Supporting experiments include HSI reconstruction and representation probes, SAR kNN probes, component ablations, and resolution/tiling controls. The authors report 57.55 validation COCO mAP on SARDet-100K versus 55.60 for their DenoDet V2 reproduction, while acknowledging that their 55.41 test mAP falls below the cited DenoDet V2 result of 56.39. Their ablations suggest modest representation-transfer benefits, with target-scale preservation contributing more substantially.

Critical strengths
  1. A practically meaningful formulation of unpaired cross-modal transfer. The method uses HSI supervision only where spatial correspondence is available within an HSI image, while a separate SAR consistency objective adapts the shared student without requiring paired HSI–SAR scenes. HSI and both teachers are discarded before inference (pp. 3–5). This is consequential because it addresses a realistic constraint on multimodal learning: auxiliary sensing data can support training without becoming a deployment requirement. It is the paper’s strongest conceptual contribution.

  2. Explicit investigation of the sources of detection improvement. Beyond evaluating four SAR benchmarks, the paper includes representation probes and controls for initialization, loss components, resolution, and tiling (Tables 3, 5, S7, S11–S12). The authors acknowledge that representation gains are modest and target-scale preservation has a larger effect. This strengthens the work’s potential scientific value by helping distinguish the contribution of representation learning from detector configuration, although the reporting inconsistencies identified elsewhere limit how confidently these controls can currently be interpreted.

Critical weaknesses
  1. Internal inconsistencies undermine confidence in the central experimental evidence. The complete five-metric SARBench result for YAMAL-C is labeled as validation performance in Table S11 but appears identically as test performance in Table 4 (pp. 8, 22). Additionally, averaging the six SARDet-100K class AP values in Table S9 gives 54.88 mAP, inconsistent with 54.51 in Table 4 (pp. 8, 21). These discrepancies leave the correspondence between splits, checkpoints, and reported results unresolved. Their reconciliation is essential before the claimed improvements can support acceptance.

  2. The evidence does not establish a consequential benefit from the central HSI-transfer mechanism. Removing all HSI alignment changes detection mAP only from 58.71 to 58.61, while improving the representation probe’s macro-F1 from 0.438 to 0.469 (Table S7, p. 19). Without repeated runs, this small detection difference does not establish a reliable contribution. Moreover, HSI reconstruction results do not directly demonstrate improved SAR noise robustness, and clean-only versus corruption-trained teacher controls are missing. This is decisive because HSI-derived robustness transfer is the paper’s principal learning contribution, whereas resolution and inference aggregation account for substantial reported improvements.

Questions for authors
  1. Which splits and checkpoints produced Tables 4, S9, and S11? Why does the complete SARBench metric vector appear as both validation and test performance, and why do the SARDet-100K class APs average to 54.88 rather than the reported 54.51? Can corrected results be traced to their evaluation records?

  2. On SAR-AIRcraft, do S and E use identical trained weights, initialization, training data, and resize settings, differing only in inference augmentation? If so, what are the individual-view results underlying the 13.44-point mAP improvement; if not, which changes contribute to that difference?

  3. Does the HSI teacher process individual spectra or spatial–spectral blocks during distillation, and how are spatially corresponding targets extracted? What are the exact definitions and application points of the initialization-anchor and VICReg losses?

1 comment
Reviewer_dEZnOpenReview ↗

My score is remained unchanged.

As the authors did not response. My concerns have not been addressed, therefore my score is remained unchanged.

Reviewer_oHb4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

SAR target detection is constrained by costly annotation, by speckle, and by small targets. This paper presents SSDenoDet, which treats recovery from corrupted input as a transferable prior. Hyperspectral imagery is densely sampled and strongly correlated across bands, so masked-autoencoding pretrained on it yields an encoder that tolerates missing and corrupted measurements; that encoder is frozen as a teacher and its spatial feature geometry is distilled into a grayscale SAR encoder through a cross-modal dual-stream objective, while a centered exponential-moving-average teacher enforces consistency between independently speckled views of unlabeled SAR, and an initialization anchor with variance-covariance regularization guards against collapse. On four independently trained benchmarks, namely SARDet-100K, AIR-SARShip, SAR-AIRcraft, and HRSID, the transferred initialization yields a small detection gain; the best results follow when it is combined with a scale-preserving feature pyramid and high-resolution inference, and the controlled comparisons attribute most of the small-object improvement to that scale-preserving design rather than to the representation transfer itself.

Critical strengths

The paper explicitly concedes the boundary of its own claim. This transparency is rare and valuable.

Critical weaknesses

The key weakness is the motivation. The title advertises denoising, but denoising an HSI cube and denoising a SAR image are different problems, and the pretraining offers no bridge between them. In HSI the corruption is a sensor artifact: dead lines, stripes and band-wise noise distort measurements along a dense spectral axis, and the missing content can be inferred from neighbouring bands because spectra are smooth. In SAR the corruption is multiplicative speckle from coherent imaging, there is no spectral axis to interpolate from, and recovery depends on spatial statistics rather than cross-band continuity. The teacher is pretrained against the former, yet no experiment measures whether the resulting encoder responds better to speckle, and the SAR branch handles speckle with its own Gamma augmentation, so the transferred content cannot be attributed to any denoising ability, a point the paper leaves unexamined. Because the supervision assumes neither paired data nor scene correspondence, whatever transfers is at most a generic feature prior, and the paper never explains why natural-image features or SAR self-supervision would not serve as well; that the limitations section concedes optical reflectance and microwave backscatter remain physically different points the same way. The controls that would identify the factor are absent: the frozen teacher is never replaced by a natural-image or SAR-only self-supervised model, nor by a MAESSI checkpoint pretrained without structured corruption. Removing all HSI alignment gives 58.61 validation mAP against 58.71 for the full model, a difference within noise, with better macro-F1 and higher effective rank; SAR self-distillation alone falls below a static ImageNet initialization; at a matched 512 input the ImageNet initialization leads by roughly 3 mAP, and the HSI-distilled encoder overtakes it only at 800 pixels with the calibration term. The paper states that its one-factor controls do not support an independent gain claim for any single loss, and attributes the small-object improvement mainly to a scale-preserving detection design.

Questions for authors
  1. The paper's claim rests on the auxiliary modality being hyperspectral. The same CMDS recipe and detector protocol should therefore be run with the frozen teacher replaced by a natural-image MAE teacher, by a SAR-only self-supervised teacher, and by a MAESSI checkpoint pretrained without stripe and dead-line corruption. Should any of these match the reported gain, the attribution to HSI would not hold.

  2. What evidence links the transferred representation to speckle robustness? The title and abstract describe the contribution as denoising, yet no measurement connects the HSI-distilled encoder to SAR's actual degradation. Can accuracy be reported as a function of ENL, or under injected multiplicative speckle, for the HSI-distilled encoder against the ImageNet initialization?

  3. Table S7 reports 58.61 validation mAP when all HSI alignment is removed, against 58.71 for the full model, with better macro-F1 (0.469 vs 0.438) and higher effective rank (1098 vs 945), and the matched-resolution control has ImageNet initialization ahead by about 3 mAP at 512 input. On what evidence is the HSI contribution claimed, and how are these two results reconciled with it?

Reviewer_V29SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents SSDenoDet, which transfers a corruption-robust masked-autoencoding representation learned from HSI to SAR. The authors propose a training pipeline for SAR object detection that combines a frozen hyperspectral masked-autoencoder teacher, a YAMAL student, SAR self-distillation, and supervised FPN detector. The motivation is to transfer corruption robustness from HSI to SAR, but the formulation, specification, and experiments do not establish a sufficiently clear account of what is transferred or how its effectiveness is verified.

Critical strengths

Leveraging hyperspectral imaging (HSI) representations as prior information for synthetic aperture radar (SAR) object detection offers a promising research direction, particularly when paired HSI–SAR acquisitions are unavailable. Representations learned from HSI could potentially provide complementary semantic cues to guide SAR feature learning and improve object discrimination.

Critical weaknesses
  1. The manuscript lacks logical coherence and a well-organized structure. For instance, the Introduction provides insufficient background on the research topic, the Related Work section lacks a comprehensive and systematic review of existing studies, and the methodological overview is overly brief, making it difficult for readers to follow the technical details.

  2. SSDenoDet-E is not a like-for-like efficiency comparison with a single DenoDet V2 model in Table 1. The paper acknowledges the extra passes, but the table headline combines the representation-transfer claim with an ensemble gain. A clearer report and single-model performance for each benchmark, are needed to establish practical advantage.

  3. The ablation studies lack a clear and rigorous logical structure. For example, the evidence does not yet establish that HSI-derived representation transfer is the main source of detection gains. The authors report only a +0.80 mAP validation gain over ImageNet initialization on SARBench.

Questions for authors
  1. What is the motivation behind this study, and why is it necessary to leverage HIS representations to guide object detection in SAR imagery? The authors should provide a clearer explanation of the research motivation and the rationale for incorporating HSI representations into SAR-based object detection.

  2. The current experimental evidence is still insufficient to fully substantiate the claimed superiority of the proposed method over existing approaches. Can the comparison against DenoDet V2 in Table 1 be repeated under a matched inference budget?

  3. The ablation studies are largely limited to a straightforward presentation of numerical results. Can you provide more comprehensive and targeted visualization analyses rather than merely listing quantitative results.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.